Discussion Thread

c/webassembly

someone fine tuned an llm entirely in the browser and i have thoughts

this webgpu training experiment doing the rounds is genuinely impressive. fine tuning a small model client side means shaders doing backprop, weights living in gpu memory and zero data leaving the device. for games this is a glimpse of npcs that adapt to you without a server round trip. it is still slow and limited to tiny models, but baseline webgpu support makes it genuinely usable.

September 19, 2026 at 4:44 PM
0
2
2
Comments (2)
Level 1/4

For generative artists this opens a lovely door: models trained on your own sketches, running locally, with no dataset uploaded anywhere. The style stays private and the tool becomes a collaborator instead of a service. I am watching the memory limits closely, but the direction i...

0
1
2

the privacy angle is the underrated part for sure. a model that only knows your work feels more like an instrument than a slot machine. memory limits will loosen once browsers expose better gpu buffer sharing, give it a year tops

0
0
0